AMEND is an inference‑time framework for text detoxification that separates detoxification into role‑conditioned proposal generation, isolated multidimensional criticism, and constraint‑first selection. It evaluates diverse candidates with role‑blind critics, enforces strict semantic and fluency feasibility, and ranks them by utility. The lightweight AMEND Lite variant limits the search to a single targeted repair round when residual toxic spans remain, achieving higher joint scores and reducing inference requests by 28.4% compared to unconstrained multi‑round loops.
By Tianyu Wang
The paper compares data-level and training-level methods for balancing language model capability and privacy leakage on synthetic medical records with embedded canaries. Direct finetuning achieves the highest capability but leaks the most private information, while DP‑SGD reduces leakage at a significant cost to performance. Generating privacy‑aware synthetic data—especially with uniqueness‑aware rewriting—offers a better trade‑off, and more capable rewriting models tend to lower downstream privacy risks, though gains are inconsistent with increased compute.
By Zihao Zhao, Avi Schwarzschild, Casey Meehan, Katherine Lee, Zhili Feng, Dylan Sam
The paper investigates how large language models (LLMs) used as labelers for toxic content differ from human crowd labels, focusing on the bias that LLMs, particularly Claude Haiku, over‑flag second‑person address (“you”) as toxic. On a dataset of 1.25 million news comments, Haiku incorrectly flags a much higher proportion of non‑toxic comments containing “you” compared to human raters, leading to a false increase in measured toxicity with reply depth. The authors show that rewriting “you” in the third person removes most of this bias, and that a certification approach—estimating laws rather than trusting individual labelers—can produce unbiased toxicity metrics even when using LLM scores.
"Why it matters": The study reveals that relying on LLM labelers without correcting for their bias can distort our understanding of toxic discourse online, potentially leading to misguided moderation policies.
By Liang Yue, Eric A. Vance
The paper argues that many vision‑language model (VLM) hallucination metrics mistakenly treat every object mentioned in a caption as a claim that the object is present. It introduces NAME‑Bench, a dataset of 25,497 non‑assertive mentions labeled as depictive, purpose, negation, or hedged, and shows that current detectors flag these as hallucinations at high rates. Adding non‑assertive mentions as negatives improves detection on LLaVA‑1.5 but requires much more data for newer VLMs, indicating that conflating mention with assertion is a persistent problem.
By Hsin-Ling Hsu, Neeraja Kirtane, Kuan-Hao Huang
RedDarwin is an agentic framework that evolves jailbreak skills by treating reusable attack mechanisms as modular skills rather than fixed prompts. It starts with 47 base skills and autonomously revises, composes, and selects them using feedback from target models, producing 65 evolved skills that improve 44 of the original attacks. The evolved skills raise the mean attack success rate on AdvBench from 18.8% to 47.3% and achieve high success on unseen frontier LLMs, including 100% ASR on GPT‑5.6‑Sol with only ten queries.
By Zhouyuan Ma, Yutao Wu, Hanxun Huang, Xiang Zheng, Yi Liu, Xingjun Ma
The paper introduces CFC26, an expanded fish‑counting dataset and evaluation framework designed to mimic real‑world sonar deployment scenarios. It extends the earlier CFC22 dataset to eleven locations across nine Pacific salmon rivers, adding new metrics (nMNE and nMANE) that align with expert workflows and uncover biases not seen with frame‑level detection. The authors demonstrate that these metrics reveal directional counting errors that could have ecological consequences.
By Madison Van Horn, Suzanne Stathatos, Sevan Brodjian, Justin Kay, Kai Van Brunt, Erik Young, Sara Beery, Pietro Perona, Michael Hobley
The paper introduces SCR-V2A, a video-to-audio generation framework that focuses on Semantic-Critical Regions (SCRs)—localized visual cues directly related to sound events. SCR-V2A first identifies these regions via sound-related visual localization, then encodes them with an adapter-based dual visual encoder to produce a control representation that captures both global context and fine-grained sound evidence. This representation conditions an audio generation model, and experiments on in-domain and out-of-domain benchmarks show improved distribution matching and audio quality with few trainable parameters.
By Yun Fang, Wei Guo, Zhuonan Liang, Heng Wang, Weidong Cai
X‑Splat introduces a unified Gaussian splatting framework that jointly reconstructs RGB and X‑ray 3D scenes. By sharing a single set of geometric parameters across modalities and adding appearance attributes only to surface‑proximal primitives, the method avoids redundant geometry and alignment artifacts. Cross‑modal registration is achieved through clean multi‑view stereo and X‑ray geometric proxies, and a surface‑depth consistency loss guides RGB reconstruction, resulting in robust alignment, higher RGB rendering quality, and a more compact representation compared to separate models.
By Wangze Xu, Chu Zhou, Mae Yamaguchi, Linrui Dai, I-Chao Shen, Takashi Ijiri, Imari Sato
Kingfisher is a fast, model‑tuning‑free framework for personalized image generation using visual autoregressive models. It decomposes early autoregressive scale codes into content and style components conditioned on prompts, then recombines these components from two real images into a compact semantic prefix. The frozen generator refines the remaining scales, achieving strong content preservation and superior style and text alignment while cutting pair‑specific preparation time by 54× compared with diffusion methods and 7.5× compared with other VAR approaches.
By Dinh-Khoi Vo, Akihiro Sugimoto, Minh-Triet Tran, Trung-Nghia Le
The paper describes the CERTH-ITI team’s approach to the TRECVID 2025 Ad-hoc Video Search (AVS) and Video Question Answering (VQA) tracks. For AVS, a two-stage framework uses foundation vision‑language models for initial retrieval and a Multimodal‑LLM for reranking, improving contextual matching without training. In VQA, an audio‑visual MLLM is fine‑tuned and enhanced with inference‑time scaling; answer generation employs majority voting over responses from random frame subsets, while multiple‑choice responses use mean‑pooling logits, yielding top performance on VQA metrics.
By Andreas Goulas, Damianos Galanopoulos, Ioannis Patras, Vasileios Mezaris
FloodVLM-Pix is a new language‑guided pixel‑level flood understanding model introduced in the arXiv paper "FloodVLM-Pix: Language‑Guided Pixel‑Level Flood Understanding." The authors created the first large‑scale remote sensing flood multi‑modal benchmark, comprising 131,200 image patches and 443,100 textual descriptions of flood severity and evolution. FloodVLM‑Pix incorporates a flood prior module to mitigate attentional bias and a multi‑temporal flood pixel decoder with a general‑specific feature fusion mechanism to improve flood segmentation accuracy.
"whyItMatters":"The paper presents a benchmark and model that enhance both flood description quality and pixel‑level segmentation, addressing shortcomings of existing vision‑language models on flood‑related tasks."
By Chengyang Zhang, Zhihao Wang, Zhiyu Long, Xueqian Wang, Quanming Yao, Wei Wu, Gang Li
Video Diffusion Avatars (VDAs) fine‑tune video diffusion models to generate photo‑realistic, multi‑view consistent avatars from a single image, allowing precise control over camera viewpoint, facial expression, and upper‑body pose. VDAs replace traditional 3D morphable models with a learned expression space, overcoming viewpoint bias through disentanglement and enabling independent control of expression, pose, and viewpoint via proxy‑mesh renderings. The method trains efficiently—34 hours on a single A100—and outperforms state‑of‑the‑art single‑image 3D avatars, improving novel‑view reenactment metrics and producing physically plausible hair dynamics, shadows, reflections, and refractions.
By Simon Giebenhain, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Zhe Chen, Matthias Nie{\ss}ner
The study explores how to fine‑tune visual brain encoding models so that their internal representations better match individual human perceptual similarity judgments. Using similarity data from 100 images collected from eight participants in the Natural Scenes Dataset, the authors adapted a generic similarity model (DreamSim) to create personalized perceptual similarity models for each subject. They then fine‑tuned each participant’s baseline encoder with either generic or personalized similarity information, finding that models trained with personalized judgments were most aligned with human perceptual similarity and preserved inter‑subject correlations of those judgments.
By Daniel Chong, Haomiao Chen, Keith W Jamison, Raja Marjieh, Ilia Sucholutsky, Nori Jacoby, Amy Kuceyeski
GatewayBench is a new evaluation framework that systematically studies how the design of Model Context Protocol (MCP) gateways influences AI agent capability, efficiency, and security. It categorizes gateways into flattened and hierarchical architectures and tests them on 524 tasks across six domains using a pool of 11,118 real-world tools from 251 MCP servers, including task-irrelevant tools to simulate large-scale deployments. Experiments on seven widely deployed gateways reveal significant differences in task completion, tool-use efficiency, and attack resilience, showing strong interactions with the underlying model and agent harness.
By Yitian Zheng, Zhaorun Chen, Haoyu Li, Bo Li
TransNSNet is a Transformer‑enhanced multi‑scale adversarial framework designed for CT‑to‑PET image synthesis. The study curated a paired CT‑PET dataset from a public NSCLC Radiogenomics collection and evaluated the framework against several baselines, reporting that TransNSNet achieved the best mean performance in nine of ten metrics while maintaining a smaller parameter footprint. Statistical analysis found no significant superiority over other models after correction, and the authors caution that synthesized PET‑like images should not replace clinically acquired PET scans.
By Shaojun Li, Xiaoyu Deng, Ikuo Keshi
TSA: Temporal Slot Activation for Persistent Object-Centric Video Representation proposes a lightweight mechanism that learns per-slot, per-frame activation scores to control slot lifecycle in unsupervised video object-centric learning. By gating slot updates and decoder participation when objects are absent or occluded, TSA reduces state drift and reconstruction interference, improving object decomposition and temporal identity preservation. The method is evaluated on MOVi-C, MOVi-E, YTVIS-HQ, and OVIS benchmarks, showing significant gains, especially on long, heavily occluded videos.
By Duc Nguyen, Sieu Tran, Hao Vo, Khoa Vo, Duy Minh Ho Nguyen, Nghi D. Q. Bui, Anh Nguyen, Long Mai, Ngan Le
The paper introduces a new family of concentration metrics, called α-RC, based on Rényi divergence to assess how the magnitudes of Jacobian minors in multilayer perceptrons are distributed. Unlike the existing Participation Ratio, these metrics are bounded between 0 and 1 regardless of network size, with 0 meaning uniform spread and 1 indicating dominance by a single minor. For α=2, the α-RC metric reduces to a transformed version of the Participation Ratio, and experiments on single-hidden-layer MLPs demonstrate its practical advantages.
By Juan Carlos Fern\'andez, Asunci\'on Jim\'enez-Cordero, Jos\'e Francisco Paris
The paper investigates explanation fairness in hate‑speech detection, noting that high classification accuracy does not guarantee equally plausible explanations for all demographic groups. Using the HateXplain dataset, the authors introduce Plausibility‑EFG, a metric that quantifies the largest difference in explanation plausibility across groups by comparing model‑generated token attributions with human‑annotated rationales. Experiments with BERT, RoBERTa, and an adversarially trained ADV‑RoBERTa show that the latter improves both classification performance and alignment with human rationales while reducing explanation disparities, and the study highlights that predictive fairness and explanation fairness are related but distinct objectives.
By Suryansh Talukdar (Purdue University Fort Wayne), Taha Sarwat (Purdue University Fort Wayne), Jonathan Rusert (Purdue University Fort Wayne)
The paper introduces Veil-then-Reveal (VtR), a user-side framework that pseudo‑anonymizes personally identifiable information (PII) in prompts sent to large language models (LLMs) under a zero‑trust model. VtR replaces detected PII with realistic, same‑type surrogates using format‑preserving encryption for structured data and a keyed‑hash dictionary for semantic data, while preserving linguistic and structural cues. The system achieves high detection recall (97.9%), strong protection of detected values (95.2%), and near‑perfect round‑trip restoration (99.9%) across three datasets.
By Mohamed Shaaban, Mohamed Elmahallawy, Thomas E. Carroll
PreciseEdit is a new one‑step image editing framework that improves semantic transport by applying strong target‑directed changes only along reliable directions and spatial locations. It is model‑agnostic, training‑free, and inversion‑free, and uses a dual‑noise‑scale strategy to balance strong editing with background fidelity and structural consistency. Experiments on PIE‑Bench show that PreciseEdit outperforms existing methods across a range of editing tasks.
By Jingyi Wang, Xiao Lu